The AI giant has also released its much awaited ‘superapp’ in the form of ChatGPT Work.
Token efficiency takes centre stage with OpenAI’s latest family of AI models that supposedly outperform rival Anthropic’s Claude Fable 5 on some tasks.
The models were subject to a staggered launch at the request of the US government, which has taken an increasingly restrictive approach to sharing advanced home-grown AI models over concerns around cyber safety and technological dominance.
CEO Sam Altman told CNBC that the flagship version of the model – GPT-5.6 Sol – is 54pc more token-efficient on agentic coding tasks, performing “as good or better” when compared to competing models on the market.
According to the company, GPT-5.6 Sol scored 53.6 on an adaptive reasoning evaluation, beating Fable 5 by around 13 points. At medium reasoning, it beat Fable by around 11 points at roughly one-quarter of the estimated cost, OpenAI said.
Meanwhile, GPT-5.6 Terra, a more balanced version of the model, and GPT-5.6 Luna, the most cost-efficient version in the family, outperform Fable 5 at around one-sixteenth the cost, according to OpenAI.
On the Artificial Analysis Intelligence Index, Sol, at its maximum reasoning capacity, comes close to Fable 5’s ability while taking 61pc less time at roughly half the estimated cost, OpenAI said.
Skyrocketing token costs have become the newest challenge to AI adoption. Uber made headlines earlier this year for blowing through its entire AI coding budget for 2026 in just four months, while Microsoft started cancelling its developers’ Claude Code licences months after enabling them.
Improvements in token efficiency are simply unable to keep up with the industry’s massive AI adoption spree, which, in turn, has made way for more niche companies such as SambaNova to step up to address inference pitfalls.
OpenAI, which is reportedly considering handing the US government administration a 5pc stake in the $852bn company, is, however, unhappy with giving the government priority access to its latest models.
“We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders and global partners who need them,” the company said in a blogpost.
“We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks”.
Altman told CNBC that the company worked with US officials, including commerce secretary Howard Lutnick, treasury secretary Scott Bessent and national cyber director Sean Cairncross, on getting approval for its new models. He hopes the regulatory approach will be global.
“Everybody will get access,” he told the publication. “It’s not like the US is going to disproportionately benefit here.”
GPT-5.6 Sol does not cross the ‘cyber critical threshold’ under OpenAI’s preparedness framework, the company said. Despite being capable of identifying the building blocks of an exploit, the model did not autonomously produce a functional full-chain exploit under test conditions, according to its maker.
“GPT‑5.6 is trained to refuse prohibited cyber assistance, including when users attempt to disguise their intent or jailbreak the model,” the company said.
‘Superapp’ makes debut
The AI giant has also released its much awaited ‘superapp’ in the form of ChatGPT Work, an AI agent that works by gathering information across a user’s apps and workflows.
The new agent, powered with GPT-5.6, represents OpenAI’s pivot towards the more profitable enterprise market where Anthropic has gained a strong footing.
OpenAI’s coding tool Codex – with more than 5m weekly users – is being merged into ChatGPT Work, providing additional benefits to the more than 1m who use the tool for work outside software development, according to the company.
The agent can reason through multi-step tasks and create materials that follow templates and reference files. Nearly all of OpenAI’s teams, including non-coding functions such as finance and sales, use ChatGPT Work and Codex, the company said.
The new launch comes after OpenAI cut a number of products from its portfolio, including its image generation model and app Sora, and shelved plans for a version of ChatGPT that would allow users to have sexual chats with AI.
The company’s standalone Atlas AI browser is the latest offering to be shut down. Atlas was initially expected to be integrated into the new superapp.
OpenAI has also launched a new generation of voice models called GPT-Live, which, according to the company, “make[s] talking with AI feel much more like having a real conversation”.
The model is built to listen and speak at the same time, and can interject with phrases such as ‘mhmm’ or ‘yeah’ to show that it is listening, according to OpenAI.
Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.














You must be logged in to post a comment Login